ReAct: Reasoning and Acting

By Satwik · April 7, 2026

ReAct is a prompting paradigm, published in 2022 and hugely influential through 2023, that interleaves reasoning traces with actions. Instead of forcing a model to either think out loud or take an action, ReAct has it alternate: produce a thought, take an action such as a search or lookup, observe the result, reason again, and continue until it answers. The reasoning steps help plan and adjust, while the actions ground the process in external information, reducing pure hallucination by letting the model check facts against tools.

Its importance is foundational to the year's agents. Most of the frameworks and autonomous systems covered in these notes run on some variant of the ReAct loop, and its thought-action-observation structure became the mental model for how an LLM agent operates. It made the interplay between internal reasoning and external tool use concrete and easy to implement with prompting alone.

The security relevance follows from that ubiquity. Because the observation step feeds tool output back into the model's context as text, ReAct-style loops are a canonical prompt-injection surface: content returned by a search or a page can contain instructions that the model treats as part of its reasoning stream and acts on. The visible reasoning trace is a double-edged asset -- it aids debugging and auditing, but it can also be manipulated or can leak intent, and a plausible-looking trace does not guarantee the underlying action was sound. Understanding ReAct is essentially prerequisite to reasoning about any 2023-era agent, because the observe-and-reason step is where untrusted data meets a model that is about to act.